Skip to content

training_inference..md

Latest commit

 

History

History
38 lines (27 loc) · 1.45 KB

training_inference..md

File metadata and controls

38 lines (27 loc) · 1.45 KB

Train & inference

Train SparseOccVLA

Train the sparse occupancy encoder first:

bash tools/dist_train.sh projects/configs/OmniDrive/sparseoccvla_stage1_4d_600q.py 8 --work-dir work_dirs/stage1_4d_600q/

Then fine-tune the entire model end-to-end, without any additional operations:

bash tools/dist_train.sh projects/configs/OmniDrive/sparseoccvla_stage3_4d_600q_forecasting.py 8 --work-dir work_dirs/stage3_4d_600q_forecasting/

Evaluation SparseOccVLA

1. Scene Understanding

bash tools/dist_test.sh projects/configs/SparseOccVLA/sparseoccvla_stage3_4d_600q_forecasting.py ckpts/stage3_4d_600q_forecasting.pth 8 --save_path results/stage3_4d_600q_forecasting

cd evaluation
python eval_language.py ../results/stage3_4d_600q_forecasting

2. Occupancy Forecasting

python tools/test.py projects/configs/SparseOccVLA/sparseoccvla_stage3_4d_600q_forecasting.py ckpts/stage3_4d_600q_forecasting.pth --eval_occ

(Optional)Train SparseOccVLA-Lidar

We also explore using raw point clouds as a form of weak supervision, which removes the reliance on dense semantic occupancy labels and improves the practical applicability of SparseOccVLA.

bash tools/dist_train.sh projects/configs/SparseOccVLA/sparseoccvla_stage1_3d_600q_lidar.py 8 --work-dir work_dirs/stage1_3d_600q_lidar/

bash tools/dist_train.sh projects/configs/SparseOccVLA/sparseoccvla_stage3_4d_600q_forecasting.py 8 --work-dir work_dirs/stage3_3d_600q_lidar/